Minimum Viable Exposure: A Data-Minimisation Standard for What AI Models Should See

AI inference pipelines routinely assemble far more personal data into a model’s context than the task requires: retrieval, memory and tool outputs pull in identities, account histories and sensitive records by default. Utility-preserving data minimisation for language-model inputs has recently been formalised for user-authored prompts; what is still missing is a reporting standard that a deployment can apply to the context it assembles itself, and that auditors can compare across releases. We propose Minimum Viable Exposure (MVE), a measurement and reporting standard for informational least privilege at inference time. MVE defines a sensitivity-weighted exposure function, an Exposure Ratio (ER) between the exposure actually sent and the minimal exposure that preserves task utility, and an MVE Score in [0, 100]. We contribute a six-class weight table mapped to GDPR categories (including an explicit weight for pseudonymised data), a greedy ablation protocol whose estimate is conservative by construction (it upper-bounds MVE and therefore never overstates over-exposure), a five-stage enforcement pattern (Context Customs), and a reporting template. In a synthetic financial-assistant example, minimisation removes 92.9% of tokens (token ratio 14.1), while the sensitivity-weighted Exposure Ratio is 203.2 (MVE Score 0.49); we decompose this gap and show that pseudonymising direct identifiers alone barely changes it. MVE is intended to complement GDPR Articles 5(1)(c) and 25 and the HIPAA minimum-necessary standard; empirical validation is left to a subsequent version.